AlexandrAI AGENTS.md
agents-md

LMSSEP Neuro-Symbolic Reasoner MCP Server Agent Guide

Project-specific AGENTS.md guidance for the LMSSEP Neuro-Symbolic Reasoner MCP server, symbolic state workflow, deterministic action selection, planning, execution simulation, and decoding.

AGENTS.md

Guidance for coding agents working on LMSSEP Neuro-Symbolic Reasoner MCP Server. This server models task-oriented reasoning as explicit symbolic state plus neural-language encoding and decoding.

Project overview

The server registers a multi-step Neuro-Symbolic Reasoner toolchain: nsr_initialize_session, nsr_encode, nsr_update_state, nsr_reasoning_engine, nsr_planner, nsr_execute, and nsr_decode. Use it for workflows that need typed state, policy enforcement, audit trails, and deterministic next-action selection.

Setup commands

bashcd <project-root>/genesis/LMSSEP/neuro_symbolic_reasoner
pip install -r requirements.txt
python main.py

Configure runtime through environment variables and configs/config.py. Generated session state belongs under output directories, not source.

Tool workflow

  1. Initialize with a natural-language task description and structured task definition.
  2. Encode user input into symbolic state.
  3. Update state with extracted parameters, constraints, and facts.
  4. Run deterministic symbolic reasoning to compute the next action.
  5. Plan executable steps from that action.
  6. Execute or simulate execution, then update state.
  7. Decode the final symbolic state into a natural-language response.

Code map

  • src/tools/neuro_symbolic_reasoner.py owns state management, reasoning history, tool registration, planning instructions, execution simulation, decoding, and session file output.
  • configs/config.py defines server metadata and output directories.
  • src/utils/session_persistence.py supports persisted sessions.

Testing and verification

Use a synthetic task definition with intents, required parameters, policies, and tools. Verify invalid JSON is rejected, required parameters are carried into state, policy rules drive the same action for the same state, planner output references available tools, and decode stays faithful to state.

Safety and output hygiene

The reasoning engine should be deterministic, but generated plans and simulated execution still need review. Never publish task definitions containing private API names, customer data, credentials, or real endpoint details. Redact generated session JSON before sharing.

Commit and PR rules

State schema, policy logic, and output file changes need explicit compatibility notes. Include one complete sanitized lifecycle example when changing any registered tool contract.